v1.1.4

Fixed

• A panic in any instruction handler can no longer crash the agent pod. The
  inbound-instruction dispatch (go handleInstruction) had no recover boundary
  anywhere in the binary, so a single handler panic (a control-plane payload
  parse, a client-go / SQL / serialization edge, or any future handler) would
  unwind the goroutine and CrashLoopBackOff the whole per-cluster control surface
  (uploads, webhooks, builds, SQL, Harbor). Dispatch now goes through
  safeHandleInstruction, which recovers, logs the value + stack, replies with an
  error for the instruction's tag (so the caller is not left hanging), and lets
  the stream keep serving. Mirrors the node agent's existing guard.